Monte Caro simulations

Transcription

1 Monte Caro simulations Monte Carlo methods - based on random numbers Stanislav Ulam s terminology - his uncle frequented the Casino in Monte Carlo Random (pseudo random) number generator on the computer Less glamorous than roulette tables or cards, but faster... >10 9 random numbers per second Monte Carlo simulations in statistical physics normally refers to importance sampling of configurations (e.g., spins) generating configurations with probability equal to the Boltzmann probability MC simulations show clearly how phase transitions can happen when N 30

2 Monte Carlo simulation of the Ising model The Metropolis algorithm [Metropolis, Rusenbluth, Rosenbluth, Teller, and Teller, Phys. Rev. 1953] Generate a series of configurations (Markov chain); C1 C2 C3 C4... Cn+1 obtained by modifying (updating) Cn changes satisfy the detailed-balance principle P change (A B) P change (B A) = W (B) W (A) = e E(A)/T W (A) Starting from any configuration, such a stochastic process leads to configurations distributed according to W the process has to be ergodic - any configuration reachable in principle it takes some time to reach equilibrium Metropolis algorithm for the Ising model. For each update perform: select a spin i at random; consider flipping it σi -σi compute the ratio R=W(σ1,...-σi,...,σN)/W(σ1,...σi,...,σN) - for this we need only the spins neighboring i generate random number 0<r 1; accept flip if r<r (go back to old config else) P change (A B) =P select (B A)P accept (B A) P select =1/N, P accept = min[w (B)/W (A), 1] These probabilities satisfy detailed balance 31

3 Symmetry breaking (magnetic phase transition) for h=0 A magnetized state, <m> 0, breaks a symmetry (E invariant under all σi -σi) strictly, mathematically we must have <m>=0 symmetry breaking (phase transition) can take place when N how can we understand the symmetry breaking for N large but finite? Time series of simulation data; magnetization vs simulation time for T<Tc M N = m = 1 N N i=1 σ i There is a characteristic reversal time between m>0 and m<0 configurations reversal time diverges for N the symmetry can be broken on practical time scales for finite (large) N also mechanism of phase transitions in real magnets (and other systems) 32

4 Another way to look at it: magnetization distribution probability distrubution (histogram) of m during the simulation m = 1 N N i=1 σ i single-peak distribution for T>Tc double-peak distribution for T<Tc peaks become sharper for increasing N no probability to fluctuate between m<0 and m>0 peaks for N - have to go through low-probability m 0 configurations Why this peak structure? balance between large number of m 0 configurations with high energy small number of m 1 configuration with low energy entropy dominates at hight T, internal energy at low T F = E ST 33

5 Binder ratios and cumulants Consider the dimensionless ratio R 2 = m4 m 2 2 We can compute R2 exactly for N for T<Tc: P(m) δ(m-m*)+δ(m+m*) m*= peak m-value R2 1 for T>Tc: P(m) exp[-m 2 /a(n)] a(n) N -1 R2 3 (properties of Gaussian integrals) The Binder cumulant is defined as (n-component order parameter; n=1 for Ising) U 2 = 3 ( n +1 n ) { 1, T < R Tc 2 0, T > T c 2D Ising model; MC results Curves for different L normally cross each other close to Tc Extrapolate crossing for sizes L and 2L to infinite size converges faster than single-size Tc defs. 34

6 Computing expectation values and their statistical errors Definition: Monte Carlo sweep = N spin-flip attempts a natural unit of simulation time measure observables after every (or every n) sweep Boltzmann probability accounted for at sampling stage Q = 1 N s N s i=1 Q i, N s = number of samples is the estimate for the true expectation value; Q Q, (N s ) Statistical errors (error bars): Q = Q ± σ Q the measurements are not statistically independent independent only after a number of sweeps >> autocorrelation time Divide the simulation into B bins, M sweeps in each bin; Ns=BM bin averages: Qb,b=1,..., B Q = 1 B Q b, σ 2 1 B Q = ( B B(B 1) Q b Q) 2 b=1 If M is sufficiently large (>> autocorrelation time) the average and error are statistically sounds (corresonding to independent Gaussian-distributed data) probability of true value being inside the error bars 2/3 b=1 35

7 Autocorrelation functions characterization of how measurements become statistically independent A Q (t) = Q(i + t)q(i) Q 2 Q 2 Q 2, ( e t/θ,t ) the autocorrelation time Θ grows as T Tc (diverges for N, T Tc) T/J=3.0 > Tc! T/J = = Tc! This problem can be largely overcome by using cluster algorithms for standard Ising, XY, Heisenberg,... but not in all cases, e.g., in the presence of external fields, frustrated systems,... 36